second-brain-rag

Built this to figure out if you can give an AI agent the same kind of structured memory a human uses in Obsidian — markdown files, wikilinks, a knowledge graph — instead of just dumping everything into a vector database.

Turns out you can, and it works better for stable curated knowledge (policy docs, SOPs, underwriting rules) because the agent follows [[wikilinks]] to pull connected concepts into context automatically.

---

How it works

flowchart TD
    Docs(["📄 Source Documents\n4 x insurance .md files"])

    subgraph Ingest["① INGEST"]
        I1["Parse YAML frontmatter\ntitle · type · tags · links"]
        I2["Extract concept nodes\nMistral-7B via Bedrock"]
        I3["Build wikilink edges\ncross-doc + same-doc"]
        I1 --> I2 --> I3
    end

    subgraph Route["② ROUTE — SLM Classifier"]
        R1["IN_CORPUS\nkeyword match → direct load"]
        R2["HYBRID_RETRIEVAL\nBM25 + BGE-M3 + RRF"]
    end

    subgraph Retrieve["③ RETRIEVE"]
        B["BM25\nkeyword overlap score"]
        S["BGE-M3\nsemantic embeddings · cosine sim"]
        F["RRF Fusion\n1÷(rank+60) combined score"]
        B --> F
        S --> F
    end

    subgraph Enrich["④ ENRICH"]
        W["Follow wikilinks\nup to 2 hops from top results"]
        W2["Pull connected concept bodies\ninto LLM context"]
        W --> W2
    end

    subgraph Generate["⑤ GENERATE"]
        C["Claude Haiku\ngrounded answer · sources attached"]
    end

    Docs --> Ingest
    Ingest --> Route
    Route --> R1 --> Enrich
    Route --> R2 --> Retrieve --> Enrich
    Enrich --> Generate

    style Ingest fill:#0d1a12,stroke:#00e5a0,color:#00e5a0
    style Route fill:#0d1520,stroke:#00c2ff,color:#00c2ff
    style Retrieve fill:#1a1408,stroke:#f5a623,color:#f5a623
    style Enrich fill:#180d24,stroke:#9b6dff,color:#9b6dff
    style Generate fill:#0d1a12,stroke:#00e5a0,color:#00e5a0

Loading

---

What it does

Four insurance documents get ingested into an in-memory knowledge graph. Each document becomes a set of typed concept nodes connected by wikilinks. When you query it, a small router model (Mistral-7B) decides whether to do a direct corpus lookup or kick off hybrid retrieval — BM25 + BGE-M3 semantic embeddings fused with RRF. Top results get enriched by following their wikilinks up to 2 hops. Then Claude generates the answer.

Three pages

• / — explains the full idea: Tiago Forte → Obsidian → Karpathy → Google OKF → this
• /pipeline — live ingestion with the knowledge graph building in real time, plus a query interface
• /graph — view the current corpus graph, filter by concept type, click nodes for details

The OKF format

Each source document is plain markdown with YAML frontmatter:

---
title: Loss Adjustor Appointment Threshold
type: claims_procedure
source: claims_handling_sop
tags: [claims, loss-adjustor, threshold]
links:
  - [[underwriting_guidelines--property-valuation-methods]]
  - [[claims_handling_sop--total-loss-declaration]]
---

Same format as Obsidian internally. The agent reads what a human would read — wikilinks and all.

Stack

• AWS Bedrock (Mistral-7B for routing, Claude Haiku for generation)
• BGE-M3 via sentence-transformers
• BM25 via rank-bm25
• RRF fusion (custom, ~15 lines)
• FastAPI + SSE for the live ingestion stream
• D3 v7 for the force-directed graph

Run it

pip install -r requirements.txt
cp .env.example .env  # add your AWS keys
python serve.py
# http://localhost:8001

Hit Ingest Source Documents first, then query.

The idea

Andrej Karpathy put it well: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase." This is that, with an insurance underwriting corpus as the wiki.

Google formalised the same format as OKF (Open Knowledge Format) — markdown + YAML frontmatter + wikilinks + a type field. We implement the same spec from scratch, no Google tooling needed at this scale.

License

This project is licensed under the MIT License. See the LICENSE file for details.